Technical practice
How would you design a system to predict delivery dates at different points in the shopping journey?
What to coverThe posting identifies delivery-date prediction under changing network conditions as a team challenge. Suggested preparation: discuss what information is available at each prediction point, how to avoid using future information, and how you would balance accuracy with customer trust and operational KPIs.
How would you use SQL, Python, and visualization to find useful training features?
What to coverExploratory data analysis and proficiency with these tools are stated in the posting. Suggested preparation: explain how you would examine missing data, changing patterns, and feature relationships, then check whether promising features remain available when the model makes predictions.
How would you evaluate the quality of delivery-date predictions over time?
What to coverThe role includes defining success metrics and regularly assessing model outputs. Suggested preparation: distinguish prediction error from the effects of early or late estimates on customer trust, and explain how you would inspect performance as the delivery network changes.
What would you consider when taking a large-scale ML model from research to production?
What to coverThe posting calls for model design, deployment, and maintenance, with collaboration from engineering partners. Suggested preparation: cover reproducible training, version control, deployment checks, ongoing monitoring, and the trade-off between model complexity and operational reliability.
How might a graph neural network help detect policy abuse despite identity manipulation?
What to coverGraph-neural-network-based policy-abuse detection is listed as a team project. Suggested preparation: explain what relationships a graph could represent, why connected evidence might complement individual records, and how you would evaluate false positives and robustness without assuming the team's implementation.
How would you estimate which customer-service policy is most effective for a particular customer?
What to coverThe posting names uplift modeling for customer-service policy personalization and lists causal inference as preferred familiarity. Suggested preparation: distinguish predicting an outcome from estimating a policy's incremental effect, and discuss comparison groups, selection bias, and evaluation.
How would you assess a vision-LLM system for product-refund diagnostics and resolution?
What to coverThe posting lists visual diagnostics and resolution for product refunds as a team project, while hands-on LLM experience is a preference. Suggested preparation: discuss how you would assess diagnostic accuracy, unsupported outputs, and the consequences of an incorrect resolution; avoid assuming a particular model or workflow.
When would you consider a multi-armed bandit rather than a fixed policy comparison?
What to coverMulti-armed bandits are listed as preferred familiarity, and the role involves balancing short- and long-term business value. Suggested preparation: explain exploration versus exploitation, when adaptive allocation could be useful, and why measuring longer-term effects may complicate the choice.
How would you make a model-training and evaluation workflow reproducible?
What to coverVersion-control practices are required; feature stores, MLflow, orchestration tools, cloud platforms, and Docker are listed as nice-to-haves. Suggested preparation: describe how you would track data and model versions, rerun evaluations, and choose workflow components based on the problem rather than assuming Wayfair uses a specific stack.
How would you decide whether a more complex model delivers enough business value?
What to coverThe role calls for large-scale models that maximize ROI while balancing short- and long-term value. Suggested preparation: compare a simpler baseline with a more complex approach using output quality, maintenance effort, operational costs, and effects on the stated customer and business goals.
Behavioural practice
Tell me about a time you defined an ML problem with product or commercial stakeholders.
What to coverThe posting emphasizes cross-functional problem definition and iterative solutions. Suggested preparation: use a real situation, explain how you translated stakeholder goals into a tractable question and revised the approach, then state the result.
Describe a time you worked with engineers to deploy and maintain a model.
What to coverEngineering collaboration and delivery from conception through production are stated expectations. Suggested preparation: describe the situation, your specific contribution to a deployment or maintenance decision, and the observable result.
Tell me about a time data analysis changed your modeling plan.
What to coverThe role includes uncovering patterns, relationships, and training features through exploratory analysis. Suggested preparation: give a real situation, the analysis and decision you made, and what happened afterward.
Describe a time you explained a model-quality concern to a nontechnical audience.
What to coverThe posting requires communication with technical and nontechnical audiences and regular assessment of model-output quality. Suggested preparation: outline the situation, how you conveyed the concern and its trade-offs, and the resulting decision or action.
Tell me about a time you balanced an immediate result against a longer-term ML outcome.
What to coverBalancing short- and long-term business value is an explicit responsibility. Suggested preparation: use a real situation, explain the options and your action, and report the result without claiming an impact you could not verify.